pstack, rebuilt for Salesforce. Rigorous agent-engineering skills tuned to the realities of the Salesforce platform: governor limits, multi-tenancy, metadata-driven deployments, the 75% Apex test gate, scratch orgs, and the sf CLI.
sfstack is a conversion of Lauren Tan's pstack (MIT) - the skill set behind her ~2,000 PRs/month - re-designed from first principles for a managed, multi-tenant cloud. It is grounded in Salesforce Well-Architected (Trusted, Easy, Adaptable) and the official Salesforce developer documentation. Every Salesforce-specific rule in these skills cites the platform guidance it comes from.
Companion site with the full research, mapping, and interactive walkthroughs: https://sfstack.vercel.app
pstack's two hardest prerequisites already ship with the Salesforce platform:
- The control CLI exists. pstack Part 1 teaches you to build a small CLI so agents can drive your app like a user. On Salesforce, the sf CLI already launches (scratch orgs), drives (data, Apex), observes (debug logs, test results), and tears down. What is missing is the verification skill that wires it together per project - sfstack generates one.
- Disposable parallel environments exist. pstack runs fleets of cloud agents; Salesforce has scratch orgs - ephemeral, source-tracked orgs. Every parallel agent gets its own throwaway org. (Mind the DevHub daily creation caps - the platform governor-limits the harness too.)
What genuinely changes in a governed, multi-tenant runtime:
- Governor limits are uncatchable. Breach one and the transaction dies with
System.LimitException- no try/catch saves you. Agents must prevent breaches by design: bulkify by default, write selective queries, favor async for volume. (Execution Governors and Limits) - Tests are a deployment gate, not a nicety. Production deploys require 75% org-wide Apex coverage, and
sf project deploy validatemakes a check-only deploy + test run the pre-merge proof. (Writing Tests, project deploy validate) - Metadata is the code. Flows, validation rules, and objects are XML source. Well-Architected is explicit: don't use org-based development; adopt source-driven development with the Salesforce CLI. (Resilient - ALM)
- You can't run the app locally. Verification happens against real orgs: deploy to a scratch org, run Apex tests, drive the Lightning UI through a browser session, capture debug logs as evidence.
52 skills (21 core + 8 Salesforce-native principles + 23 upstream pstack principles vendored in) + 10 playbooks. Self-contained: install sfstack alone and you get everything - no separate pstack install needed.
| skill | what it does |
|---|---|
/sf-mode |
Sticky entry mode. Routes your task to a playbook, applies the principles, keeps rigor on across turns. Start here. |
/sf-setup |
Verifies sf CLI, DevHub auth, default scratch org definition; writes project config. Run once per project. |
/sf-create-verification-skill |
Meta-skill: interviews your sfdx project and generates a project-local verification skill (launch / doctor / drive / evidence / cleanup) plus a metadata-aware Feature Map. The centerpiece. |
/sf-maintain-verification-skill |
Periodic pass keeping the verification skill and Feature Map honest as the org's source changes. |
/sf-architect |
Five-phase design loop: ground in the real schema and automations, sketch data model + sharing model first, arena competing designs, implement, scrap when the signals say the design is wrong. |
/sf-arena |
N competing designs, each in its own scratch org; evidence picks the winner. |
/sf-swarm |
N parallel workers across scratch orgs or metadata slices; one aggregated report. |
/sf-interrogate |
Multi-lens adversarial review: bulk safety, CRUD/FLS + sharing, SOQL selectivity, transaction/async design, test quality, and the Well-Architected pattern/anti-pattern checklists. |
/sf-blast-radius |
What else a change could break, via the metadata dependency graph (Tooling API) - with the safety fact proven by running tests, not asserted. |
/sf-tdd |
Failing Apex test first, then the fix. Real transaction semantics: Test.startTest/stopTest, data factories, no SeeAllData. |
/sf-deploy |
Validate-first shipping: check-only deploy with tests, then deploy, keep deployment/test-run IDs as evidence, post-deploy smoke check. |
/sf-how |
Traces runtime mechanics the Salesforce way: order of execution, automation map, schema, debug logs. |
/sf-why |
Investigates intent: git history, tickets, docs, setup audit trail. |
/sf-teach |
Runs how + why, weaves one plain explanation. |
/sf-recall |
Rebuilds your working context from history and the repo record. |
/sf-reflect |
Reviews a landed task and routes learnings into concrete skill edits. The self-improvement loop. |
/sf-show-me-your-work |
Reviewable decision trail (TSV): what, why, evidence - deployment IDs and test-run IDs included. |
/sf-figure-it-out |
Designs an auditable one-off playbook when none fits. |
/sf-technical-writing |
Diátaxis + Google developer style doc standard. |
/sf-unslop |
Cuts AI tells from writing. |
/sf-bro |
Restates the last message in plain language. |
| principle | rule | grounded in |
|---|---|---|
principle-bulkify-by-default |
All data operations against collections; never single-record paths. Bulkification is necessary but not sufficient for large data volumes. | Well-Architected: Reliable - Performance |
principle-selective-queries |
Positive operators, no LIKE wildcards, only needed fields, no LIMIT 1, no ALL ROWS. |
Well-Architected: Reliable - Performance |
principle-respect-the-shared-runtime |
Governor limits are uncatchable. Prevent by design; probe with the Limits class. |
Apex Governor Limits |
principle-test-real-transactions |
Test.startTest/stopTest, data factories, no SeeAllData, assert outcomes. 75% is a floor, not a target; proof means it ran in a real org. |
Apex Testing |
principle-metadata-is-code |
Flows, validation rules, objects live in version control. Source-driven, never org-based, development. | Well-Architected: Adaptable - Resilient |
principle-enforce-access-explicitly |
CRUD/FLS and sharing are deliberate: with sharing, stripInaccessible, no silent escalation. |
Well-Architected: Trusted - Secure |
principle-async-for-volume |
Favor queueable/batch for volume; Bulk API 2.0 for LDV; mind latency tradeoffs. | Well-Architected: Reliable - Performance |
principle-no-hardcoded-ids |
No hard-coded record, user, or record-type IDs in Apex or Flow. | Well-Architected: Easy - Automated |
All 23 of Lauren Tan's engineering principles ship inside this repo under skills/principle-*, vendored from pstack @ f5bdd68 and expanded with Salesforce application notes (MIT, (c) 2026 Lauren Tan - see NOTICE.pstack). /sf-mode indexes them next to the Salesforce-native ones and reads them from the local files.
attack-the-premise · boundary-discipline · build-the-lever · encode-lessons-in-structure · exhaust-the-design-space · experience-first · fix-root-causes · foundational-thinking · guard-the-context-window · laziness-protocol · make-operations-idempotent · migrate-callers-then-delete-legacy-apis · minimize-reader-load · model-the-domain · never-block-on-the-human · outcome-oriented-execution · prove-it-works · redesign-from-first-principles · separate-before-serializing-shared-state · sequence-verifiable-units · subtract-before-you-add · test-behavior-not-implementation · type-system-discipline
investigation · bug-fix · feature · perf-issue · deploy · data-migration · flow-change · multi-phase-plan · autonomous-run · opening-a-pr
Works with any tool that supports the Agent Skills spec.
# Claude Code, Cursor, Codex, OpenCode, ...
npx skills add krika2810/sfstack
# Claude Code plugin install (alternative)
/plugin marketplace add krika2810/sfstack
/plugin install sfstack@sfstackThen, inside any sfdx project:
/sf-setup
/sf-create-verification-skill
/sf-mode <your first real task>
All 47 upstream pstack skills are accounted for: 20 core skills ported and adapted, 4 deliberately not ported (typescript-best-practices → replaced by the 8 Salesforce-native principles; automate-me, make-bot-ui, no-comments → tool-specific or work as-is upstream), and the 23 upstream engineering principles are vendored in this repo and expanded with Salesforce application notes (MIT, see NOTICE.pstack) - /sf-mode indexes them alongside the Salesforce-native ones and reads them from the local files. The 23 upstream playbooks are represented by 10 Salesforce playbooks covering the platform's actual task shapes. Full interactive mapping table: https://sfstack.vercel.app
- Lauren Tan (@poteto) - pstack and the two-part guide (Part 1, Part 2). This project is a conversion of her ideas, not affiliated with her or Cursor. pstack ships MIT.
- Salesforce - sf-skills (official curated agent skills: complementary - sfstack is the discipline/verification layer, sf-skills covers generation workflows), Well-Architected, and the Salesforce developer documentation cited throughout.